Model result, rank 24 of 27

NVIDIA Nemotron 3 Ultra

OpenRouter · nvidia/nemotron-3-ultra-550b-a55b · extra-high reasoning. Public result card with the model’s overall score, lane measurements, runtime and cost telemetry, and the ranking formula.

Compare all models Download public JSON

Figure 1 Lane profile against the cohort Each bar is one measured lane on the shared 0 to 100 scale. The thin mark is the cohort median; the bright mark is the best score any ranked model reached on that lane.

Full / Agenticlane rank #26 of 27

67.15

SWE MVPlane rank #22 of 27

58.63

Hard Intelligencelane rank #23 of 27

64.56

NVIDIA Nemotron 3 Ultra
Cohort median
Cohort best

Overall score63.45

Rank #24

Full / Agentic67.15

Full rank #26

SWE MVP58.63

SWE rank #22

Hard Intelligence64.56

Hard rank #23

Measured cost$0.489

100.0% reliability

Overall

All-around publication view

Score63.45

Formulamean(Full, SWE, Hard Intelligence)

BasisOpenRouter · nvidia/nemotron-3-ultra-550b-a55b · extra-high reasoning

The overall score averages the measured major lanes while keeping each source measurement visible.

Lane 01

Full / Agentic benchmark

Final67.15

Capability97.06

Agentic87.84

Pass rate94.3%

Prompts43

This lane captures instruction following, structured behavior, tool discipline, and general agentic reliability.

Lane 02

Software engineering MVP

SWE score58.63

Focused final59.30

Capability58.63

Daily driver60.63

Prompts24

This lane is closer to implementation usefulness: source handling, architecture cleanliness, and deliverable quality.

Lane 03

Hard Intelligence diagnostic

Hard score64.56

Active inquiry93.08

Online adaptation8.75

Self-repair56.40

Authority integrity100.00

Hard Intelligence measures active inquiry, online adaptation, evidence-driven self-repair, and authority/salience integrity.

Telemetry

Runtime economics

Total cost$0.489

Cost / scored item$0.0065

Seconds / timed item70.71s

Runtime coverage100.0%

Recorded tokens / item8.8k

Token coverage100.0%

Cost, time, and token basis are normalized telemetry. They explain tradeoffs; they do not overwrite the capability score yet.

Interpretation

Why the result lands here.

The model is stronger in the Full/Agentic lane than in the SWE lane; the overall score is therefore shown with both component lanes visible. Hard Intelligence score is 64.56 and contributes to the overall score alongside Full/Agentic and SWE. Hard Intelligence is published alongside Full/Agentic and SWE for the current ranking. Full long-context probe rescored under suite v4 (rubric: correct answer zeroed by the phrase list).